Abstract: Streaming video data presents unique challenges for AI systems, particularly in detecting and segmenting events in real time without extensive labeled datasets. Traditional supervised approaches often struggle with scalability and adaptability, making self-supervised learning an attractive alternative for video understanding.
In this talk, Dr. Saket Anand and Dr. Sudeep Sarkar will introduce a perceptual prediction framework designed to tackle these challenges. The framework leverages self-supervised techniques to learn meaningful representations directly from raw video streams, enabling robust event detection and segmentation without reliance on exhaustive manual annotation.
Key themes will include:
The role of perceptual prediction in modeling temporal dynamics of streaming video.
How self-supervised learning can reduce dependence on labeled data while improving adaptability.
Applications in surveillance, autonomous systems, and multimedia analysis.
Challenges of deploying such frameworks in real-world environments, including efficiency and scalability.
The presentation will highlight both theoretical foundations and practical demonstrations, showing how perceptual prediction can advance the state of the art in video analytics. By combining self-supervision with event-focused modeling, this approach opens new opportunities for intelligent systems to interpret complex, continuous visual data streams in real time.